Diffulex is a Diffusion Language Model Serving Engine built on PagedAttention-style runtime primitives. It provides a unified runtime for KV cache management, block scheduling, prefix reuse, MoE execution, CUDA graph replay, HTTP serving, and model-specific diffusion samplers.
Diffulex is also the runtime engine behind the Multi-Block Diffusion Language
Models (MBD-LMs) line of work. Native Block Diffusion LMs perform
Single-Block Diffusion (SingleBD): one noisy block per forward pass,
creating a store bubble where no new output is produced. Multi-Block
Diffusion (MultiBD) removes this bottleneck with a bounded running-set of
consecutive blocks, enabling decode-store overlap and inter-block parallelism.
MBD-LMs are BD-LMs post-trained with Multi-block Teacher Forcing (MultiTF)
to handle practical MultiBD states, and Diffulex executes them with an optimized
Block Buffer runtime that preserves static shapes for CUDA Graph replay.
In the engine, this is exposed as decoding_strategy=multi_bd.
The README is intentionally brief. Use the documentation for installation, configuration, benchmarks, serving, and development notes:
| Goal | Go to |
|---|---|
| Read the full documentation | Diffulex Documentation |
Install Python, CUDA, and the tested vllm==0.23.0 dependency |
Installation |
| Run the first LLaDA2-mini command | Quickstart |
| Check supported models and strategy combinations | Models |
| Tune runtime and YAML parameters | Configuration |
| Run GSM8K and other benchmark workflows | Benchmark |
| Start HTTP serving and the local demo visualization | Server |
| Use Diffulex as a research backend | Research Engine |
| Add a model, decoding strategy, or kernel | Developer Guide |
For reproducing the MBD-LMs experiments, use the Diffulex
mbd-lms branch
(CUDA 12).
For engine development, open-source contributions, or exploring new decoding
algorithms and turning them into runnable systems, use the
main branch. main
contains ongoing runtime and model-specific optimizations, so its behavior and
performance profile may differ from the experiment reproduction branch.
The main branch requires CUDA 13.
Diffulex main is built for researchers who want a real backend for block-level
dLLM inference ideas. Its Block Buffer backend separates logical block state,
running-set/cache policy, and paged KV plus Triton kernel execution, so new
SingleBD, MultiBD, TokenMerge, edit, uniform diffusion, or cache-oriented
algorithms can be implemented as bounded strategy changes instead of full
serving-system rewrites.
Diffulex focuses on cache-aware block-wise dLLM decoding through a single core backend that supports multiple main strategies: MultiBD (SingleBD at BufSz=1, full MultiBD at BufSz≥2), Token Merge + Edit (DMax), Edit Sampling / T2T (LLaDA2.1), D2F MultiBD, Fast-dLLM-v2 Dual Cache, and DiffusionGemma. The most complex part of adding a strategy is modifying the request state machine — sampler-only strategies like DMax require no state machine changes at all.
Supported model families include Dream/DiffuCoder-style dense dLLMs, Dream reasoner, Stable-DiffCoder, LLaDA, Fast-dLLM-v2, SDAR, SDAR-MoE, LLaDA2, and DiffusionGemma. See the Models documentation for the up-to-date compatibility matrix.
For questions, development discussion, and collaboration, join the Discord or our WeChat group:
We would like to thank Nano-vLLM, vLLM, mini-sglang, SGLang, and dInfer, whose designs informed parts of Diffulex's early backend, paged attention, serving architecture, and dLLM inference optimizations. Diffulex is developed by the DENG Lab at Shanghai Jiao Tong University.